AnchorFormer

This repository includes the pure code of AnchorFormer model in the paper.

UPDATE

2023-07-28: update a pretrained model on PCN.
2023-06-15: update model files to support arbitrary output points.

Environment

These model are implemented in PyTorch (1.8.1+cu102) version.

The version of operation system in docker is Ubuntu 18.04.6 LTS.

The python version is 3.8.13.

The GPU is NVIDIA Tesla V100 (16GB) and the CUDA version is CUDA 10.2.

Install

Requirements

pip install -r requirements.txt

c++ extensions

bash ./extensions/install.sh

Standard PointNet++ lib (ref to "https://github.com/erikwijmans/Pointnet2_PyTorch")

Usage

train

pretrained PCN model PCN results

CUDA_VISIBLE_DEVICES=${GPUS} python -m torch.distributed.launch --master_port=${PORT} --nproc_per_node=${NGPUS} main.py --launcher pytorch --sync_bn ${PY_ARGS}

example:

train from start

CUDA_VISIBLE_DEVICES=0,1,2,3 WORLD_SIZE=4 python -m torch.distributed.launch --node_rank=0 --nnodes=1 --master_port=13232 --nproc_per_node=4 main.py --launcher pytorch --sync_bn --config ./cfgs/PCN_models/AnchorFormer.yaml --exp_name try_to_train_anchorformer (--val_freq 10 --val_interval 50)

resume from last break-point

CUDA_VISIBLE_DEVICES=0,1,2,3 WORLD_SIZE=4 python -m torch.distributed.launch --node_rank=0 --nnodes=1 --master_port=13232 --nproc_per_node=4 main.py --launcher pytorch --sync_bn --config ./cfgs/PCN_models/AnchorFormer.yaml --exp_name try_to_train_anchorformer --resume

resume from specified break-point

CUDA_VISIBLE_DEVICES=0,1,2,3 WORLD_SIZE=4 python -m torch.distributed.launch --node_rank=0 --nnodes=1 --master_port=13232 --nproc_per_node=4 main.py --launcher pytorch --sync_bn --config ./cfgs/PCN_models/AnchorFormer.yaml --exp_name try_to_train_anchorformer --start_ckpts {CKPT_PATH}.pth

test

CUDA_VISIBLE_DEVICES=${GPUS} python main.py --test ${PY_ARGS}